JP7414907B2 - Pre-trained model determination method, determination device, electronic equipment, and storage medium - Google Patents

Description

The present disclosure relates to the field of artificial intelligence technology, specifically to the field of computer vision and deep learning technology, and can be applied to scenes such as image processing and image recognition, and in particular, a method and device for determining a pre-trained model. , related to electronic devices and storage media.

Pre-trained models have been widely applied to improve the effectiveness of upper-level artificial intelligence tasks. In upstream tasks, pre-trained models are pre-trained with a large amount of training data, and in downstream tasks, models can be trained with a small amount of training data. Good prediction results can be obtained just by training. Therefore, it is extremely important how to reduce the training cost and improve the training efficiency of pre-trained models.

The present disclosure provides a method for determining a pre-trained model, a determination device thereof, an electronic device, and a storage medium.

In a first aspect of the present disclosure, the steps include: obtaining a plurality of candidate models; performing structure coding based on model structures of the plurality of candidate models to obtain a structure code of each of the candidate models; mapping the structural code of each candidate model using a predefined encoder to obtain a frequency domain code of each candidate model; and determining the model performance of each candidate model based on the frequency domain code of each candidate model. and determining a target model from a plurality of candidate models as a pre-trained model based on model performance parameters of each candidate model. be done.

In a second aspect of the present disclosure, an acquisition module configured to acquire a plurality of candidate models, and performing structural coding based on model structures of the plurality of candidate models, generate a structure code of each of the candidate models. a mapping module configured to map a structural code of each said candidate model using a trained encoder to obtain a frequency domain code of each said candidate model; a prediction module configured to predict model performance parameters of each said candidate model based on frequency domain codes of said candidate models; and a prediction module configured to predict model performance parameters of each said candidate model based on frequency domain codes of said candidate models; a determination module configured to determine a target model from the candidate models.

A third aspect of the present disclosure includes at least one processor and a memory communicatively connected to the at least one processor, the memory storing instructions executable by the at least one processor. and the instructions are executed by the at least one processor, an electronic device is provided, wherein the at least one processor is capable of performing the pre-trained model determination method described above.

In a fourth aspect of the disclosure, there is provided a non-transitory computer-readable storage medium having computer instructions stored thereon for causing a computer to perform the pre-trained model determination method described above.

In a fifth aspect of the disclosure, a computer program product is provided which, when executed by a processor, implements the method for determining a pre-trained model described above.

Note that the content described in this summary section does not specify essential or important features of the embodiments of the present disclosure, nor does it limit the scope of the present disclosure. Other features of the disclosure will become more apparent from the following specification.
The drawings are for a better understanding of the solution and do not limit the disclosure.

1 is a schematic flowchart of a method for determining a pre-trained model according to a first example of the present disclosure. 2 is a schematic flowchart of a method for determining a pre-trained model according to a second example of the present disclosure. 12 is a schematic flowchart of a method for determining a pre-trained model according to a third example of the present disclosure. FIG. 7 is a schematic configuration diagram of a pre-trained model determining device according to a fourth example of the present disclosure. FIG. 7 is a schematic configuration diagram of a pre-trained model determining device according to a fifth example of the present disclosure. FIG. 2 is a block diagram of an electronic device for implementing a method for determining a pre-trained model according to an embodiment of the present disclosure.

Exemplary embodiments of the present disclosure are described below with reference to the drawings, in which various details of embodiments of the present disclosure are included to aid in understanding and are provided by way of illustration only. should be considered. Accordingly, those skilled in the art may make various substitutions and modifications to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for the sake of clarity and brevity, the following description omits descriptions of well-known features and structures.

Currently, pre-trained models are widely applied to improve the effectiveness of upper-level artificial intelligence tasks, and by pre-training the pre-trained model with a large amount of training data in the upstream task, a small amount of training data is required in the downstream task. Preferable prediction results can be obtained simply by training the model. Therefore, how to reduce the training cost and improve the training efficiency of pre-trained models is of great importance.

The present disclosure provides a method for determining a pre-trained model in order to reduce the training cost of the pre-trained model and improve training efficiency. In this method, after obtaining multiple candidate models, structural coding is performed based on the model structure of the multiple candidate models to obtain the structural code of each candidate model, and then a trained encoder is used to generate the structure code of each candidate model. Mapping the structure code to obtain the frequency domain code of each candidate model, predicting model performance parameters of each candidate model based on the frequency domain code of each candidate model, and predicting model performance parameters of each candidate model based on the model performance parameters of each candidate model. Then, a target model is determined from multiple candidate models as pre-trained models. This reduces the subsequent training cost of training pre-trained models and improves training efficiency by determining the target model from multiple candidate models as pre-trained models based on the frequency domain codes of multiple candidate models. can be improved.

Hereinafter, a method for determining a pre-trained model, a pre-trained model apparatus, an electronic device, a non-transitory computer-readable storage medium, and a computer program product according to embodiments of the present disclosure will be described with reference to the drawings.

First, a method for determining a pre-trained model according to the present disclosure will be described in detail with reference to FIG.

FIG. 1 is a schematic flowchart of a method for determining a pre-trained model according to a first embodiment of the present disclosure.

Here, the entity that executes the pre-trained model determining method according to the embodiment of the present disclosure is a pre-trained model determining device, and hereinafter the determining device will be referred to as the determining device. The determining device may be an electronic device or may be located within the electronic device. This realizes determining a target model from multiple candidate models as a pre-trained model based on the frequency domain codes of multiple candidate models, reducing the subsequent training cost of training the pre-trained model, Training efficiency can be improved. In the embodiments of the present disclosure, a case will be described in which the determination device is placed in an electronic device.

Here, an electronic device is any fixed or mobile computing device capable of performing data processing, for example, a mobile computing device such as a laptop, a smartphone, a wearable device, or a fixed computing device such as a desktop computer. , or a server, or other type of computing device, but is not limited by this disclosure.

As shown in FIG. 1, the method for determining a pre-trained model may include the following steps 101-105.

In step 101, a plurality of candidate models are obtained.

Each candidate model is a combination of multiple trained submodels. The plurality of trained sub-models may be neural network models or other types of models, but are not limited in this disclosure.

In step 102, structure coding is performed based on the model structures of a plurality of candidate models to obtain a structure code for each candidate model.

In an exemplary embodiment, structure coding is performed on each of the plurality of candidate models based on the model structure of the candidate model to obtain a structure code for each candidate model.

In the structural code of the candidate model, each item corresponds to one layer of the candidate model, where one layer can be understood as one of multiple submodels that make up the candidate model, and each item corresponds to The taken value is the model type of the submodel of the layer corresponding to the item.

For example, assuming that each sub-model making up a candidate model is selected from a model set, the model set contains 10000 types of sub-models, and candidate model A has a total of 6 layers, each layer is This corresponds to one item of the structure code of candidate model A. Correspondingly, the structure code of candidate model A contains six items, each item containing 10,000 possible values. The model type of the first layer submodel of candidate model A is number 5 in the model set, the model type of the second layer submodel is number 2 in the model set, and the model type of the third layer submodel is 5. The number in the model set is 9, the model type of the 4th layer submodel is number 8 in the model set, the model type of the 5th layer submodel is number 7 in the model set, the 6th layer submodel Assuming that the model type number in the model set is 4, structure coding is performed based on the model structure of candidate model A, and the structure code of candidate model A is [5, 2, 9, 8, 7, 4] can be obtained.

In step 103, a trained encoder is used to map the structural code of each candidate model to obtain a frequency domain code of each said candidate model.

In an exemplary embodiment, a training encoder can be pre-trained such that the input of the encoder is a structural code and the output is a corresponding frequency-domain code, thus applying the structural code of each candidate model to the trained encoder respectively. By inputting the frequency domain code corresponding to the structure code of each candidate model, it is possible to obtain a frequency domain code corresponding to the structure code of each candidate model, thereby realizing mapping of the structure code of each candidate model to the corresponding frequency domain code.

At step 104, predict model performance parameters for each candidate model based on the frequency domain code of each candidate model.

Model performance parameters can characterize the performance of candidate models. The model performance parameters can include parameters representing the accuracy of the candidate model, parameters representing the processing speed of the candidate model, and the like.

In an exemplary embodiment, a correlation function that describes the correlation between a frequency domain code and a model performance parameter of a corresponding candidate model may be pre-statistically obtained, and the parameters of the correlation function may be calculated using the maximum likelihood in the frequency domain. It can be obtained based on estimation. Thereby, after obtaining the frequency domain code of each candidate model, model performance parameters of each candidate model are determined based on a correlation function that describes the correlation between the frequency domain code and the model performance parameter of the candidate model corresponding to the frequency domain code. Can be predicted. For a detailed method of obtaining a correlation function through statistics, please refer to the correlation technique, and a detailed explanation will be omitted here.

In step 105, a target model is determined from the plurality of candidate models as pre-trained models based on the model performance parameters of each candidate model.

The number of pre-trained models determined from the plurality of candidate models can be set in advance as necessary, for example, it can be set in advance to one or more, and the present disclosure is not limited thereto.

In the illustrative example, after predicting and obtaining model performance parameters for each candidate model, each candidate model is sorted in descending order of performance based on the model performance parameters, and from the multiple candidate models, the top A set number of target models can be determined as pre-trained models, and furthermore, by training the pre-trained models, the pre-trained models can be used for various tasks such as face recognition, image processing, product classification, etc. can be made suitable.

After obtaining multiple candidate models, a target model is determined from the multiple candidate models as a pre-trained model based on the frequency domain codes of the multiple candidate models, thereby eliminating the need to subsequently train each candidate model. , only the determined pre-trained model needs to be trained. Thereby, the training cost of training a pre-trained model can be reduced and training efficiency can be improved. Also, by selecting a pre-trained model based on the model performance parameters of each candidate model, from each candidate model, the candidate model with the fastest processing speed when the accuracy is the same is selected as the pre-trained model. Can be done. Furthermore, after training a pre-trained model, when performing tasks such as image processing and image recognition, you can increase the processing speed or image recognition speed of the model on specific hardware, or use low-cost hardware. can achieve the same speed and accuracy as high-cost hardware; alternatively, from each candidate model, the candidate model with the highest accuracy under the same speed conditions can be selected as the pre-trained model; After training a pre-trained model, when performing tasks such as image processing and image recognition, the accuracy of the model can be improved under comparable hardware conditions.

In the method for determining a pre-trained model according to the embodiment of the present disclosure, after obtaining a plurality of candidate models, structure coding is performed based on the model structures of the plurality of candidate models to obtain a structure code of each candidate model. , further map the structure code of each candidate model using a trained encoder to obtain the frequency domain code of each said candidate model, and calculate the model performance parameters of each candidate model based on the frequency domain code of each candidate model. predict and determine a target model from the plurality of candidate models as pre-trained models based on model performance parameters of each candidate model. This reduces the subsequent training cost of training pre-trained models and improves training efficiency by determining the target model from multiple candidate models as pre-trained models based on the frequency domain codes of multiple candidate models. can be improved.

As can be seen from the above analysis, embodiments of the present disclosure pre-train a training encoder and map the structural code of each candidate model using the trained encoder to obtain the frequency domain code of each said candidate model. be able to. Hereinafter, the encoder training process in the pre-trained model determination method according to the present disclosure will be further described with reference to FIG. 2.

FIG. 2 is a schematic flowchart of a method for determining a pre-trained model according to a second embodiment of the present disclosure. As shown in FIG. 2, the method for determining a pre-trained model may include the following steps 201-208.

In step 201, a sample structure code configured as a training sample is input to an encoder to obtain a predicted frequency domain code output by the encoder.

The sample structure code can be obtained by performing structure coding on the sample model based on the model structure of the sample model. For the process of performing structural coding on a sample model, please refer to the description of the above embodiment, and the description will be omitted here.

In step 202, the predicted frequency domain code is input to a decoder.

In step 203, the encoder and decoder are trained based on the differences between the decoder output and the sample structure code.

The encoder and decoder may each be neural network models or other types of models, but are not limited in this disclosure. The input of the encoder is a structured code and the output is a frequency domain code corresponding to the structured code, and the input of the decoder is a frequency domain code and the output is a structured code corresponding to the frequency domain code.

In an exemplary embodiment, the encoder and decoder may be trained in a deep learning manner, for example. Compared to other machine learning methods, deep learning performs better on big data sets.

When training an encoder and decoder using a deep learning method, first, one or more sample structure codes in the training samples are input to the encoder, and the prediction corresponding to the sample structure code output by the encoder is input. Obtain a frequency domain code, then input the predicted frequency domain code output by the encoder as input to a decoder, obtain the predicted structure code corresponding to the predicted frequency domain code output by the decoder, and According to the structure code, obtain the difference between the output of the decoder and the sample structure code, adjust the parameters of the encoder and decoder based on the difference between the output of the decoder and the sample structure code, and generate the adjusted encoder and Get decoder.

Then, input one or more sample structure codes in the training data to the adjusted encoder, and obtain a predicted frequency domain code corresponding to the sample structure code output by the adjusted encoder. , then take the predicted frequency domain code output by the adjusted encoder as input, input it to the adjusted decoder, and obtain the predicted structure code corresponding to the predicted frequency domain code output by the adjusted decoder. , obtain the difference between the adjusted decoder output and the sample structure code according to the sample structure code, and calculate the adjusted encoder and the sample structure code based on the difference between the adjusted decoder output and the sample structure code. Adjust the parameters of the adjusted decoder to obtain further adjusted encoders and decoders.

This allows the encoder and decoder to be trained repeatedly by constantly adjusting the encoder and decoder parameters, and when the accuracy of the predicted structure code output by the decoder meets a preset threshold, the training is terminated and the trained Obtain an encoder and trained decoder.

Through the above process, a trained encoder and a trained decoder can be obtained, the trained encoder can map the structure code of a specific model to the frequency domain code, and the trained decoder can map the frequency domain code of the specific model. could be mapped to a structural code, thereby laying the foundation for subsequently mapping the structural code of each candidate model using a trained encoder to obtain the frequency domain code of each said candidate model.

In step 204, a plurality of candidate models are obtained.

In step 205, structural coding is performed based on the model structures of a plurality of candidate models to obtain a structural code for each candidate model.

At step 206, a trained encoder is used to map the structural code of each candidate model to obtain a frequency domain code of each candidate model.

In an exemplary embodiment, after training the encoder and decoder using the training process described above, obtaining a plurality of candidate models and obtaining the structural code of each candidate model, the trained encoder is used to Structural codes may be mapped to obtain frequency domain codes for each of the candidate models.

In step 207, model performance parameters for each candidate model are predicted based on the frequency domain code of each candidate model.

Note that in the embodiments of the present disclosure, when mapping the structural code of each candidate model to the corresponding frequency domain code, the structural code can be mapped to at least a two-dimensional frequency domain code, and the at least two-dimensional frequency domain code may include, for example, at least a time dimension and an accuracy dimension, thereby improving the accuracy of prediction when predicting model performance parameters for each candidate model based on at least two-dimensional frequency domain codes of each candidate model. can.

Correspondingly, when training the encoder and decoder, after inputting a sample structure code configured as a training sample to the encoder, the encoder performs at least two-dimensional coding, and the encoder outputs at least two-dimensional A predicted frequency-domain code of the sample structure can be obtained, and then the at least two-dimensional predicted frequency-domain code is input to a decoder, and the encoder and Train the decoder. Thereby, a trained encoder is used to map the structural code of each candidate model to obtain at least two-dimensional frequency domain codes of each candidate model, and based on the at least two-dimensional frequency domain codes of each candidate model, It is possible to predict the model performance parameters of each candidate model and improve the accuracy of prediction.

At step 208, a target model is determined from the plurality of candidate models as pre-trained models based on the model performance parameters of each candidate model.

For the specific implementation process and principle of steps 204 to 208, please refer to the description of the above embodiment, and detailed description thereof will be omitted here.

In the method for determining a pre-trained model according to an embodiment of the present disclosure, a sample structure code configured as a training sample is input to an encoder, a predicted frequency domain code outputted by the encoder is obtained, and the predicted frequency domain code is transmitted to a decoder. and trains the encoder and decoder based on the difference between the output of the decoder and the sample structure code, thereby realizing the training of the encoder and decoder. Obtain multiple candidate models, perform structural coding based on the model structure of the multiple candidate models, obtain the structural code of each candidate model, and then map the structural code of each candidate model using a trained encoder. obtain the frequency domain code of each candidate model, further predict model performance parameters of each candidate model based on the frequency domain code of each candidate model, and perform pre-training based on the model performance parameters of each candidate model. A target model can be determined from a plurality of candidate models. This reduces the subsequent training cost of training pre-trained models and improves training efficiency by determining the target model from multiple candidate models as pre-trained models based on the frequency domain codes of multiple candidate models. can be improved.

As can be seen from the above analysis, embodiments of the present disclosure predict the model performance parameters of each candidate model based on the frequency domain code of each candidate model, and further predict the model performance parameters of each candidate model in advance based on the model performance parameters of each candidate model. A target model is determined from multiple candidate models as trained models. Hereinafter, the process of predicting model performance parameters of each candidate model based on the frequency domain code of each candidate model in the pre-trained model determination method according to the present disclosure will be further described with reference to FIG. 3.

FIG. 3 is a schematic flowchart of a method for determining a pre-trained model according to a third embodiment of the present disclosure. As shown in FIG. 3, the method for determining a pre-trained model may include the following steps 301-306.

In step 301, feature extraction models in the model set are combined to obtain a plurality of candidate models.

The feature extraction model may be any model in the computer vision and image processing field that is capable of extracting image features.

In an exemplary embodiment, the model set includes a plurality of trained feature extraction models (i.e., sub-models in the above embodiments), and the plurality of feature extraction models may be neural network models; Other types of models may be used and are not limited by this disclosure. In an exemplary embodiment, multiple candidate models may be obtained by selecting and combining multiple feature extraction models from the model set in a random selection manner. Alternatively, first determine the performance of each of the multiple feature extraction models in the model set, and then select several feature extraction models with superior performance from the model set and randomly combine them to obtain multiple candidate models. You may. Alternatively, a plurality of candidate models may be obtained using other methods. In the embodiment of the present disclosure, the method of acquiring multiple candidate models is not limited.

By combining the feature extraction models in the model set, multiple highly accurate candidate models can be obtained.

In step 302, structural coding is performed based on the model structures of a plurality of candidate models to obtain a structural code for each candidate model.

In step 303, a trained encoder is used to map the structural code of each candidate model to obtain a frequency domain code of each said candidate model.

For the specific implementation process and principle of steps 302 to 303, please refer to the description of the above embodiment, and detailed description thereof will be omitted here.

At step 304, a target correlation function is determined based on the task to be performed.

Here, the task to be executed is a task that needs to be executed after the pre-trained model is trained, and may be, for example, a face recognition task or a product classification task.

In an exemplary embodiment, a correlation function corresponding to each of the tasks may be predetermined, and the correlation function corresponding to each task may be a function of the frequency domain code and the time when the corresponding candidate model performs the task. Describe the correlation with model performance parameters. The parameters of the correlation function can be obtained by maximum likelihood estimation in the frequency domain. A target correlation function corresponding to the task to be executed can be determined based on the correlation function corresponding to the task to be executed and each predetermined task.

In step 305, the frequency domain code of each candidate model is substituted into the target correlation function to obtain model performance parameters of each candidate model.

In an exemplary embodiment, the target correlation function determines the frequency-domain code of each candidate model by describing the correlation between the frequency-domain code and the model performance parameters when performing the tasks that the corresponding candidate model performs. can be substituted into the target correlation function to obtain the model performance parameters of each candidate model.

Based on the task to be performed, a target correlation function is determined, and the frequency domain code of each candidate model is substituted into each of the target correlation functions to obtain model performance parameters of each candidate model. This makes it possible to determine model performance parameters when each candidate model executes a task based on the target correlation function corresponding to the task to be executed.

In step 306, a target model is determined from the plurality of candidate models as a pre-trained model based on the model performance parameters of each candidate model.

For the specific implementation process and principle of step 306, please refer to the description of the above embodiment, and detailed description will be omitted.

In the method for determining a pre-trained model according to an embodiment of the present disclosure, first, feature extraction models in a model set are combined to obtain a plurality of candidate models, and then a structure is constructed based on the model structure of the plurality of candidate models. coding to obtain the structural code of each candidate model, further mapping the structural code of each candidate model using a trained encoder to obtain the frequency domain code of each said candidate model, and then executing Based on the task to determine the target correlation function, substitute the frequency domain code of each candidate model into the target correlation function respectively to obtain the model performance parameters of each candidate model, and Then, a target model is determined from multiple candidate models as pre-trained models. This reduces the subsequent training cost of training pre-trained models and improves training efficiency by determining the target model from multiple candidate models as pre-trained models based on the frequency domain codes of multiple candidate models. can be improved.

Hereinafter, a pre-trained model determining device according to the present disclosure will be described with reference to FIG. 4.

FIG. 4 is a structural schematic diagram of a pre-trained model determining device according to a fourth embodiment of the present disclosure.

As shown in FIG. 4, a pre-trained model determination apparatus 400 according to the present disclosure includes an acquisition module 401, a coding module 402, a mapping module 403, a prediction module 404, and a determination module 405.

Acquisition module 401 is configured to acquire multiple candidate models.
Coding module 402 is configured to perform structural coding based on the model structures of the plurality of candidate models to obtain a structural code for each candidate model.
The mapping module 403 is configured to map the structural code of each candidate model using a trained encoder to obtain a frequency domain code of each said candidate model.
Prediction module 404 is configured to predict model performance parameters for each candidate model based on the frequency domain code of each candidate model.
The determination module 405 is configured to determine a target model from the plurality of candidate models as pre-trained models based on model performance parameters of each candidate model.

Note that the pre-trained model determining device according to this embodiment can execute the pre-trained model determining method of the above embodiment. The determining device may be an electronic device or may be located within the electronic device. This realizes determining a target model from multiple candidate models as a pre-trained model based on the frequency domain codes of multiple candidate models, reducing the subsequent training cost of training the pre-trained model, Training efficiency can be improved.

An electronic device is a fixed or mobile computing device capable of performing data processing, such as a notebook computer, a smartphone, a mobile computing device such as a wearable device, or a fixed computing device such as a desktop computer, or a server, Alternatively, it may be other types of computing devices, etc., and is not limited by this disclosure.

Note that the above description of the embodiment of the pre-trained model determining method also applies to the pre-trained model determining apparatus according to the present disclosure, and detailed description thereof will be omitted here.

A pre-trained model determining device according to an embodiment of the present disclosure obtains a plurality of candidate models, and then performs structure coding based on the model structures of the plurality of candidate models to obtain a structure code of each candidate model. , further map the structure code of each candidate model using a trained encoder to obtain the frequency domain code of each said candidate model, and calculate the model performance parameters of each candidate model based on the frequency domain code of each candidate model. predict and determine a target model from the plurality of candidate models as pre-trained models based on model performance parameters of each candidate model. This reduces the subsequent training cost of training pre-trained models and improves training efficiency by determining the target model from multiple candidate models as pre-trained models based on the frequency domain codes of multiple candidate models. can be improved.

Hereinafter, a pre-trained model determining device according to the present disclosure will be described with reference to FIG. 5.

FIG. 5 is a schematic configuration diagram of a pre-trained model determining device according to a fifth example of the present disclosure.

As shown in FIG. 5, the pre-trained model determination apparatus 500 may specifically include an acquisition module 501, a coding module 502, a mapping module 503, a prediction module 504, and a determination module 505. In FIG. 5, acquisition module 501, coding module 502, mapping module 503, prediction module 504, and decision module 505 have the same functions as acquisition module 401, coding module 402, mapping module 403, prediction module 404, and decision module 405 in FIG. It has a configuration.

In an exemplary embodiment, the pre-trained model determination apparatus 500 is configured to input sample structure codes configured as training samples to an encoder to obtain predicted frequency domain codes output by the encoder. A first processing module 506 and a second processing module 507 configured to input a predicted frequency domain code to the decoder train the encoder and the decoder based on the difference between the output of the decoder and the sample structure code. A training module 508 configured to.

In an exemplary embodiment, the first processing module 506 inputs a sample structure code configured as a training sample into an encoder to perform at least two-dimensional coding and at least two-dimensional predicted frequencies output by the encoder. A processing unit configured to obtain a region code.

In an exemplary embodiment, the acquisition module 501 comprises a combination unit configured to combine the feature extraction models in the model set to obtain a plurality of candidate models.

In an exemplary embodiment, the prediction module 504 includes a determination unit configured to determine a target correlation function based on the task to be performed, and a determination unit configured to respectively substitute the frequency domain code of each candidate model into the target correlation function. , an acquisition unit configured to acquire model performance parameters of each candidate model.

Note that the above description of the embodiment of the pre-trained model determining method also applies to the pre-trained model determining apparatus according to the present disclosure, and detailed description thereof will be omitted here.

A pre-trained model determining device according to an embodiment of the present disclosure obtains a plurality of candidate models, and then performs structure coding based on the model structure of the plurality of candidate models to obtain a structure code of each candidate model. , further map the structure code of each candidate model using a trained encoder to obtain the frequency domain code of each said candidate model, and calculate the model performance parameters of each candidate model based on the frequency domain code of each candidate model. predict and determine a target model from a plurality of candidate models as a pre-trained model based on model performance parameters of each candidate model. This reduces the subsequent training cost of training pre-trained models and improves training efficiency by determining the target model from multiple candidate models as pre-trained models based on the frequency domain codes of multiple candidate models. can be improved.

According to embodiments of the present disclosure, the present disclosure provides electronic devices, readable storage media, and computer program products.

FIG. 6 depicts a schematic block diagram of an example electronic device 600 in which embodiments of the present disclosure may be implemented. Electronic equipment is intended to refer to various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic equipment may also represent various forms of mobile devices such as personal digital processors, mobile phones, smart phones, wearable devices, and other similar computing devices. The components depicted herein, their connections and relationships, and their functionality are merely examples and are not intended to limit the description herein and/or the required implementation of the present application.

As shown in FIG. 6, electronic device 600 can perform various suitable operations based on a computer program stored in read-only memory (ROM) 602 or loaded into random access memory (RAM) 603 from storage unit 608. The computer includes a calculation unit 601 that can perform various operations and processes. RAM 603 can include various programs and data necessary for operating device 600. Computing unit 601, ROM 602 and RAM 603 are connected to each other via bus 604. An input/output (I/O) interface 605 is also connected to bus 604.

An input unit 606 such as a keyboard and a mouse, an output unit 607 such as various types of monitors and speakers, a storage unit 608 such as a magnetic disk or an optical disk, and a communication unit such as a network card, modem, wireless communication receiver/transmitter, etc. 609 are connected to the I/O interface 605. Communication unit 609 allows device 600 to exchange information/data with other devices via computer networks such as the Internet and/or various telecommunications networks.

Computing unit 601 may be a variety of general purpose and/or special purpose processing components with processing and computing capabilities. Some examples of computational units 601 are central processing units (CPUs), graphics processing units (GPUs), various specialized artificial intelligence (AI) computational chips, various machine execution learning model algorithm computational units, digital signal processors. (DSP), and any suitable processor, controller, microcontroller, etc. The calculation unit 601 performs the methods and processes described above, such as the method for determining a pre-trained model. For example, in some embodiments, the pre-trained model determination method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, some or all of the computer program is loaded and/or installed on electronic device 600 via ROM 602 and/or communication unit 609. When the computer program is loaded into the RAM 603 and executed by the calculation unit 601, one or more steps of the pre-trained model determination method described above may be performed. Alternatively, in other embodiments, the computing unit 601 may be configured to perform the pre-trained model determination method by any other suitable method (eg, via firmware).

Various embodiments of the systems and techniques described above herein include digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), It can be implemented in a system-on-chip (SOC), complex programmable logic device (CPLD), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may be implemented in one or more computer programs that execute and/or interpret on a programmable system that includes at least one programmable processor. The programmable processor may be an application-specific or general-purpose programmable processor and receives data and instructions from a storage system, at least one input device, and at least one output device, and the programmable processor receives data and instructions from the storage system, at least one input device, and at least one output device. The information may be transmitted to a storage system, the at least one input device, and the at least one output device.

Program code for implementing the methods of the present application can be written in any combination of one or more programming languages. These program codes may be implemented on a general purpose computer, special purpose computer, or other programmable data processing device such that, when executed by a processor or controller, the functions/acts specified by the flowcharts and/or block diagrams are performed. processor or controller. The program code may run entirely on a machine, partially on a machine, as a standalone software package, partially on a machine, and partially on a remote machine, or completely remotely. It may be executed on a machine or a server.

In the context of this application, a machine-readable medium is capable of containing or storing a program for use by or in combination with an instruction execution system, apparatus, or device. It may be a tangible medium. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. The machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the above. Not done. More specific examples of machine-readable storage media include electrical connections based on one or more lines, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), fiber optics, portable compact disk read-only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the above.

To provide interaction with a user, the systems and techniques described herein can be implemented on a computer that includes a display device (e.g., a CRT (cathode ray tube)) for displaying information to the user. or an LCD (liquid crystal display) monitor), and a keyboard and pointing device (eg, a mouse or trackball) through which a user can provide input to the computer. Other types of devices may also provide interaction with the user, for example, the feedback provided to the user may be any form of sensing feedback (e.g., visual feedback, auditory feedback, or haptic feedback). Input from the user may be received in any format, including acoustic input, voice input, or tactile input.

The systems and techniques described herein are applicable to computing systems with back-end components (e.g., as data servers), or with middleware components (e.g., application servers), or with front-end components. (e.g., a user computer having a graphical user interface or web browser, through which the user interacts with embodiments of the systems and techniques described herein), or such back-end components; , middleware components, and front-end components. The components of the system may be interconnected by any form or medium of digital data communication (eg, a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), the Internet, and blockchain networks.

A computer system can include clients and servers. Clients and servers are generally remote from each other and typically interact via a communications network. A client and server relationship is created by computer programs running on corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also called a cloud computing server or a cloud host, and is one host product in a cloud computing service system, which is different from a traditional physical host and a VPS ("Virtual Private Server" or "VPS"). (abbreviated as ``)'' service, it solved the deficiencies of difficult management and weak business scalability. The server may be a server of a distributed system or a server in combination with a blockchain.

The present disclosure relates to the field of artificial intelligence technology, specifically to the field of computer vision and deep learning technology, and can be applied to scenes such as image processing and image recognition.

Artificial intelligence is a field that studies how computers simulate certain human thought processes and intellectual behaviors (e.g. learning, reasoning, thinking, planning, etc.), and hardware-level technology also includes software-level technology. There is also. The hardware technology of artificial intelligence generally includes sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing and other technologies, and the software technology of artificial intelligence mainly includes computer vision technology, voice recognition technology, natural language It includes several aspects such as processing technology and machine learning/deep learning, big data processing technology, knowledge graph technology, etc.

According to the technical solution of embodiments of the present disclosure, a subsequent training of the pre-trained model by determining a target model from the plurality of candidate models as the pre-trained model based on the frequency domain codes of the plurality of candidate models. can reduce training costs and improve training efficiency.

Note that steps can be rearranged, added, or deleted using the various types of flows shown in the above description. For example, each step described in this disclosure may be performed in parallel, sequentially, or in a different order, but according to the technical solution disclosed in this application. There is no limitation in this specification as long as the desired result can be achieved.

The above specific embodiments do not limit the protection scope of the present disclosure. Those skilled in the art will appreciate that various modifications, combinations, subcombinations, and substitutions may be made depending on design requirements and other factors. All modifications, equivalent replacements, improvements, etc. made within the spirit and principles of this disclosure should be included in the protection scope of this disclosure.

Claims (11)

obtaining a plurality of candidate models;
performing structural coding based on model structures of the plurality of candidate models to obtain a structural code of each of the candidate models;
mapping the structural code of each candidate model using a trained encoder to obtain a frequency domain code of each candidate model;
predicting model performance parameters for each candidate model based on the frequency domain code of each candidate model;
determining a target model from the plurality of candidate models as a pre-trained model based on model performance parameters of each candidate model;
including;
predicting model performance parameters for each candidate model based on the frequency domain code of each candidate model;
determining a target correlation function based on the task to be performed;
substituting the frequency domain code of each of the candidate models into the target correlation function to obtain model performance parameters of each of the candidate models;
How to determine pre-trained models , including : inputting a sample structure code configured as a training sample into an encoder to obtain a predicted frequency domain code output by the encoder;
inputting the predicted frequency domain code into a decoder;
training the encoder and the decoder based on differences between the output of the decoder and the sample structure code;
2. The method of determining a pre-trained model according to claim 1. inputting a sample structure code configured as a training sample into the encoder to obtain a predicted frequency domain code output by the encoder;
3. The method according to claim 2, further comprising the step of inputting the sample structure code configured as the training sample to the encoder to perform at least two-dimensional coding to obtain at least two-dimensional predicted frequency domain code output by the encoder. How to determine the pre-trained model described. The step of obtaining multiple candidate models is
The method of determining a pre-trained model according to claim 1, comprising the step of combining feature extraction models in a model set to obtain the plurality of candidate models. an acquisition module configured to acquire a plurality of candidate models;
a coding module configured to perform structural coding based on model structures of the plurality of candidate models to obtain a structural code of each of the candidate models;
a mapping module configured to map a structural code of each candidate model using a trained encoder to obtain a frequency domain code of each candidate model;
a prediction module configured to predict model performance parameters for each candidate model based on the frequency domain code of each candidate model;
a determination module configured to determine a target model from a plurality of candidate models as a pre-trained model based on model performance parameters of each candidate model;
Equipped with
The prediction module is
a determining unit configured to determine a target correlation function based on a task to perform;
an acquisition unit configured to respectively substitute frequency domain codes of each candidate model into the target correlation function to obtain model performance parameters of each candidate model;
A pre-trained model decision device comprising : a first processing module configured to input a sample structure code configured as a training sample to an encoder to obtain a predicted frequency domain code output by the encoder;
a second processing module configured to input the predicted frequency domain code to a decoder;
a training module configured to train the encoder and the decoder based on differences between the output of the decoder and the sample structure code;
6. The pre-trained model determining device according to claim 5 . the first processing module,
a processing unit configured to input the sample structure code configured as the training sample into the encoder to perform at least two-dimensional coding to obtain an at least two-dimensional predicted frequency domain code output by the encoder; 7. The pre-trained model determining device according to claim 6 . The acquisition module includes:
The device for determining a pre-trained model according to any one of claims 5 to 7 , comprising a combination unit configured to combine feature extraction models in a model set to obtain the plurality of candidate models. at least one processor;
a memory communicatively connected to the at least one processor;
Equipped with
The memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor An electronic device capable of performing the pre-trained model determination method described in Section 1. A non-transitory computer-readable storage medium having computer instructions stored thereon for causing a computer to carry out a method for determining a pre-trained model according to any one of claims 1 to 4 . A computer program, when executed by a processor, implements the method for determining a pre-trained model according to any one of claims 1 to 4 .